Content
77%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
An exceptionally actionable, well-sequenced orchestration procedure with validation and recovery paths for every failure mode. Its weaknesses are length — rationale prose and time-sensitive statistics mixed into the instructions — and a monolithic body that keeps procedures the main loop rarely needs (stack merging, workflow stage splitting) inline instead of in reference files.
Suggestions
Move the self-contained sub-procedures — "A frontend stack", the Workflow stage-splitting stages, "Adopting an open pull request", and "Every wait ends" — into files under references/ (like gotchas.md) and keep one-line pointers plus the triggering condition in SKILL.md.
Cut the rationale narratives and repo statistics (e.g. "64% of this repository's recorded token spend", "85.8% of subagent spend today") or move them to a metrics reference; they are time-sensitive and do not change what the agent does.
State the load-bearing rules (TASK-BRANCH marker, premise checks, wait terminal states) as terse imperative checklists at each step, leaving the why to at most one clause, to cut token cost without losing actionability.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body never explains concepts Claude already knows and most of its length is genuine operational detail, but there is noticeable fat: rationale prose ("what survives a compaction is what was written down, not what you were holding in context"), time-sensitive repo statistics outside any deprecated section ("64% of this repository's recorded token spend currently sits in that anonymous HEAD pool", "It is 85.8% of subagent spend today"), and long narrative passages that could be cut to imperative rules. Mostly efficient with tightening needed fits 3; it is above 2 because the verbosity is occasional digression, not pervasive padding. | 3 / 5 |
Actionability | Fully executable throughout: exact commands with expected outputs (the precondition checks, the fetch refspec, `bun install --frozen-lockfile`), precise `gh api` calls with `--jq` filters, marker formats (`TASK-BRANCH: <branch>`, `CLAIMED: <repo>#<n>`), wait-script exit codes, and a `test -e` check on every named path. This is copy-paste-ready guidance covering the common cases, matching the 5 anchor. | 5 / 5 |
Workflow Clarity | Steps 00, 0, and 1–5 are explicitly sequenced with validation checkpoints everywhere: precondition commands with expected outputs, premise verification before every brief, mechanical checks before trusting a subagent's report, the CI-started count check, and terminal states plus deadlines for every wait. Each failure mode (red check, DIRTY, BEHIND, flake, dead subagent, blocked CI) has an explicit recovery path — a clear match to the 5 anchor's feedback loops and checklists. | 5 / 5 |
Progressive Disclosure | The step structure is clear and the one bundle reference is real and well signaled ("The full table, with the reason behind each, is `references/gotchas.md`" — top six inline, full table in the file, one level deep). But the ~540-line body inlines several large, self-contained subsections — "A frontend stack", the Workflow stage-splitting procedure, "Adopting an open pull request", "Every wait ends" — that belong in reference files like gotchas.md, which is the anchor-3 pattern of content that should be separate staying inline; it reaches 4 only if most content is appropriately placed, which it is not. | 3 / 5 |
Total | 16 / 20 Passed |